深度学习之分类回归

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.animation as animation
from matplotlib import font_manager as fm

# ==========================================
# 1. 中文显示设置
# ==========================================
# 中文显示设置（自动选择可用中文字体，避免乱码）
candidate_fonts = [
    'Noto Sans CJK SC', 'Source Han Sans SC', 'WenQuanYi Zen Hei',
    'SimHei', 'Microsoft YaHei', 'PingFang SC', 'Heiti SC',
    'Arial Unicode MS', 'DejaVu Sans'
]
available_fonts = {f.name for f in fm.fontManager.ttflist}
chinese_font = next((f for f in candidate_fonts if f in available_fonts), 'DejaVu Sans')

plt.rcParams['font.sans-serif'] = [chinese_font, 'DejaVu Sans']
plt.rcParams['axes.unicode_minus'] = False

# ==========================================
# 2. 生成模拟样本数据 (三分类)
# ==========================================
np.random.seed(42)
n_samples = 60
cat = np.random.multivariate_normal([1.5, 1.5], [[0.3, 0.1], [0.1, 0.3]], n_samples)
dog = np.random.multivariate_normal([4.0, 4.0], [[0.4, 0.1], [0.1, 0.4]], n_samples)
rabbit = np.random.multivariate_normal([4.0, 1.0], [[0.3, 0], [0, 0.3]], n_samples)

X = np.vstack([cat, dog, rabbit])
y = np.array([0]*n_samples + [1]*n_samples + [2]*n_samples)
Y_onehot = np.eye(3)[y] 

# ==========================================
# 3. 全局模型参数
# ==========================================
W = None
b = None
learning_rate = 0.2

# 动画与导出配置
total_frames = 150
interval_ms = 50
save_gif = True
gif_path = 'pytorch/demo_0306_training.gif'

def reset_model():
    global W, b
    # 随机初始化 W (2特征 x 3类别), b (3类别)
    W = np.random.randn(2, 3) * 0.5
    b = np.random.randn(1, 3) * 0.1

def softmax(z):
    e_z = np.exp(z - np.max(z, axis=1, keepdims=True))
    return e_z / e_z.sum(axis=1, keepdims=True)

# ==========================================
# 4. 设置绘图布局
# ==========================================
# 增加子图宽度，以便在右侧留出空间显示参数
fig = plt.figure(figsize=(12, 8))
ax = fig.add_subplot(111)
plt.subplots_adjust(right=0.75) # 给文字留出空间

# 参数文本区域只创建一次，更新时仅改文本，减少闪烁。
param_box = fig.text(
    0.78,
    0.5,
    '',
    fontsize=11,
    verticalalignment='center',
    fontfamily=chinese_font,
    bbox=dict(boxstyle='round', facecolor='wheat', alpha=0.2),
)

x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1
y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, 0.05), np.arange(y_min, y_max, 0.05))
grid_points = np.c_[xx.ravel(), yy.ravel()]

# ==========================================
# 5. 动画逻辑
# ==========================================
def update(frame):
    global W, b
    
    if frame == 0:
        reset_model()
    
    ax.clear()
    
    # --- 训练步骤 (梯度下降) ---
    logits = np.dot(X, W) + b
    probs = softmax(logits)
    error = probs - Y_onehot
    dW = np.dot(X.T, error) / X.shape[0]
    db = np.mean(error, axis=0)
    
    W -= learning_rate * dW
    b -= learning_rate * db
    
    # --- 几何绘制 ---
    Z_logits = np.dot(grid_points, W) + b
    Z_class = np.argmax(Z_logits, axis=1).reshape(xx.shape)
    
    ax.contourf(xx, yy, Z_class, alpha=0.3, cmap=plt.cm.RdYlBu)
    ax.scatter(cat[:, 0], cat[:, 1], c='red', edgecolors='k', s=40, label='猫')
    ax.scatter(dog[:, 0], dog[:, 1], c='green', edgecolors='k', s=40, label='狗')
    ax.scatter(rabbit[:, 0], rabbit[:, 1], c='blue', edgecolors='k', s=40, label='兔')
    
    loss = -np.mean(np.sum(Y_onehot * np.log(probs + 1e-9), axis=1))
    
    # --- 实时参数显示 ---
    param_text = (
        f"轮次: {frame}\n"
        f"损失: {loss:.4f}\n\n"
        f"--- 权重矩阵 W ---\n"
        f"(控制旋转角度)\n"
        f"猫: [{W[0,0]:.2f}, {W[1,0]:.2f}]\n"
        f"狗: [{W[0,1]:.2f}, {W[1,1]:.2f}]\n"
        f"兔: [{W[0,2]:.2f}, {W[1,2]:.2f}]\n\n"
        f"--- 偏置向量 b ---\n"
        f"(控制平移位置)\n"
        f"猫: {b[0,0]:.2f}\n"
        f"狗: {b[0,1]:.2f}\n"
        f"兔: {b[0,2]:.2f}"
    )
    # 在绘图区外侧显示文字（复用已有文本对象，中文显示更稳定）
    param_box.set_text(param_text)

    ax.set_title("模型训练几何演化与参数监控", fontsize=14, fontfamily=chinese_font)
    ax.legend(loc='upper left', prop={'family': chinese_font, 'size': 10})
    ax.set_xlim(x_min, x_max)
    ax.set_ylim(y_min, y_max)
    ax.grid(True, linestyle='--', alpha=0.3)

# ==========================================
# 6. 启动动画
# ==========================================
ani = animation.FuncAnimation(
    fig,
    update,
    frames=total_frames,
    interval=interval_ms,
    repeat=True,
    repeat_delay=1500,
)

# 支持导出 GIF。PillowWriter 生成的 GIF 默认可循环播放。
if save_gif:
    fps = max(1, int(1000 / interval_ms))
    print(f'使用字体: {chinese_font}')
    print(f'开始保存 GIF: {gif_path}')
    ani.save(gif_path, writer=animation.PillowWriter(fps=fps), dpi=120)
    print(f'GIF 保存完成: {gif_path}')

plt.show()